An underground structure hydraulic load numerical calculation method and system based on artificial intelligence and multi-source data

Abstract

The application discloses a kind of underground structure hydraulic load numerical calculation method and system based on artificial intelligence and multi-source data, it is related to underground engineering numerical calculation technical field, including: the multi-source data acquisition of underground structure is carried out, and index system is constructed;Multi-source data is preprocessed to obtain standard data set;Improved APNN model is constructed and trained by standard data set;Real-time data is input into improved APNN model to obtain underground structure hydraulic load numerical prediction result.The application fuses multi-source data and improved APNN, dynamically adapts geological condition change;Through teacher-student model framework, the physical consistency of APNN prediction result is improved;Historical data weighted sampling is combined with approximate field transfer learning, and the generalization ability of model under new geological scene is improved;While retaining the self-adaptive characteristics of APNN, the prediction deviation problem of pure data-driven model under complex hydrological conditions is solved by physical constraint correction.

Description

Technical Field

[0001] This invention relates to the field of numerical calculation technology for underground engineering, and more specifically to a method and system for numerical calculation of hydraulic loads on underground structures based on artificial intelligence and multi-source data. Background Technology

[0002] In porous water-bearing rocks, the calculation of hydraulic loads on underground structures is the core of engineering safety control. Existing technologies have revealed the correlation between hydraulic loads and pore water pressure and surrounding rock porosity through analytical analysis and numerical simulation, but they have obvious limitations: multi-source data, such as field measurements, physical simulations, and geological survey data, are scattered and insufficiently integrated, failing to form a dynamic perception mechanism; traditional models, such as BP neural networks, lack adaptability and are unable to cope with dynamic changes in geological conditions, such as sudden changes in pore water pressure; and the prediction results do not fully incorporate the physical constraints of hydraulic loads, such as the positive correlation between hydraulic loads and pore water pressure, resulting in large errors under extreme conditions.

[0003] Therefore, how to propose a numerical calculation method and system for hydraulic loads of underground structures based on artificial intelligence and multi-source data, and how to achieve high-precision and adaptive hydraulic load prediction in order to address the shortcomings of existing methods such as insufficient multi-source data fusion, lack of dynamic adaptability of models, and lack of physical constraints, is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] In view of this, the present invention provides a numerical calculation method and system for hydraulic loads of underground structures based on artificial intelligence and multi-source data, which solves the defects of existing methods such as insufficient multi-source data fusion, lack of dynamic adaptability of models, and lack of physical constraints, and achieves high-precision, adaptive hydraulic load prediction. To achieve the above objectives, the present invention adopts the following technical solution: A numerical calculation method for hydraulic loads on underground structures based on artificial intelligence and multi-source data includes: Conduct multi-source data collection on underground structures and construct an indicator system; The multi-source data is preprocessed to obtain a standard dataset; An improved APNN model was built and trained using a standard dataset; The real-time data is input into the improved APNN model to obtain the numerical prediction results of the hydraulic load on the underground structure.

[0005] Optionally, the indicator system includes: static geological indicators and dynamic monitoring indicators.

[0006] Optionally, the static geological indicators include: rock type, initial porosity, and lining material parameters, and the dynamic monitoring indicators include real-time pore water pressure, surrounding rock strain, and seepage flow.

[0007] Optionally, the data preprocessing includes: Outlier detection: Outliers are identified using the interquartile range method; Missing value handling: Geological zoning interpolation method is used to fill in missing values ​​in offline sensor data; Standardization: Eliminate dimensional differences through Z-score standardization.

[0008] Optionally, the improved APNN model includes: Input layer: Receives static geological indicators and dynamic monitoring indicators; Model layer: Embedding physical constraint terms, the positive correlation between hydraulic load and pore water pressure is incorporated into the Gaussian function, and the corrected output is: Where λ is the physical constraint coefficient; Adaptive layer: When the pore water pressure change rate > k, the mean vector m of the model layer is automatically updated. i and standard deviation σ i .

[0009] Optionally, the adaptive layer further includes: The model input is monitored in real time, and the input data of the indicator system is sampled at high frequency so that the model can perceive changes in the hydraulic load state of the underground structure in a timely manner. When a significant change in dynamic indicators is detected, an adaptive adjustment mechanism is triggered. The model automatically adjusts the parameters according to the real-time data to adapt to the new data distribution and characteristics, and performs online learning of the model to continuously adapt to the dynamic changes in the hydraulic load state of the underground structure and the evolution of the hydraulic load state of the underground structure.

[0010] Optionally, the improved APNN model training includes: dividing 70% of the data into a training set and 30% into a test set, and using dynamic learning rate gradient descent to optimize the parameters.

[0011] Optionally, it also includes performing model validation, which includes verifying mean squared error and physical plausibility.

[0012] Optionally, a numerical calculation system for hydraulic loads on underground structures based on artificial intelligence and multi-source data includes: Data Acquisition Module: Used for multi-source data acquisition of underground structures and construction of an indicator system; Preprocessing module: used to preprocess the multi-source data to obtain a standard dataset; Model building module: Used to build and train improved APNN models using standard datasets; Prediction module: Used to input real-time data into the improved APNN model to obtain numerical prediction results of hydraulic loads on underground structures.

[0013] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a method and system for numerical calculation of hydraulic loads on underground structures based on artificial intelligence and multi-source data, which has the following beneficial effects: This invention proposes a numerical calculation method for hydraulic loads on underground structures based on artificial intelligence and multi-source data. The method includes: collecting multi-source data on underground structures and constructing an index system; preprocessing the multi-source data to obtain a standard dataset; constructing and training an improved APNN model using the standard dataset; and inputting real-time data into the improved APNN model to obtain numerical prediction results of the hydraulic loads on the underground structures. This invention integrates multi-source data with an improved APNN model, enabling dynamic adaptation to changes in geological conditions and improving computational efficiency. Furthermore, the teacher-student model framework enhances the physical consistency of the APNN prediction results. By combining historical data weighted sampling with approximate domain transfer learning, the model's generalization ability in new geological scenarios is improved. While retaining the adaptive characteristics of APNN, physical constraint correction solves the prediction bias problem of pure data-driven models under complex hydrological conditions, reducing errors in extreme operating conditions. The system integrates a closed loop of "collection-analysis-prediction," providing a scientific basis for the safety design of underground engineering. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0015] Figure 1 This invention provides a flowchart illustrating a numerical calculation method for hydraulic loads on underground structures based on artificial intelligence and multi-source data.

[0016] Figure 2 The present invention provides a structural framework diagram of a numerical calculation system for hydraulic loads on underground structures based on artificial intelligence and multi-source data. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] This invention discloses a numerical calculation method for hydraulic loads on underground structures based on artificial intelligence and multi-source data, such as... Figure 1 As shown, it includes: Conduct multi-source data collection on underground structures and construct an indicator system; The multi-source data is preprocessed to obtain a standard dataset; An improved APNN model was built and trained using a standard dataset; The real-time data is input into the improved APNN model to obtain the numerical prediction results of the hydraulic load on the underground structure.

[0019] Furthermore, the indicator system includes: static geological indicators and dynamic monitoring indicators.

[0020] Furthermore, the static geological indicators include: rock type, initial porosity, and lining material parameters, and the dynamic monitoring indicators include real-time pore water pressure, surrounding rock strain, and seepage flow.

[0021] Furthermore, it also includes: building a distributed sensor network, deploying water pressure sensors, strain gauges, and seepage flow monitors at key sections of underground structures (such as the outer edge of the lining and deep parts of the surrounding rock), using high-frequency sampling (10Hz) to capture transient changes, and ensuring the spatiotemporal consistency of multi-source data through time synchronization technology.

[0022] In a specific implementation, the specific steps for building the distributed sensor network and controlling the spatiotemporal consistency of multi-source data include: I. Sensor Network Layout Planning Steps (1) Delineation of key monitoring areas: Based on the hydraulic load distribution characteristics of underground structures (such as the lining arch, sidewalls, bottom slab, and the area within 5 times the tunnel diameter in the surrounding rock), the monitoring areas are delineated using the hydraulic load gradient method: 1) Calculate the hydraulic load gradient coefficient , Pore ​​water pressure, This represents the distance along the structural axis; 2) When When defined as a high gradient region, the sensor deployment density is 2 sensors per meter; when At that time, it is a low gradient region with a density of 1 per meter.

[0023] (2) Sensor selection criteria

[0024] 1) Water pressure sensor: range 0-10MPa, accuracy ±0.5%FS, sampling frequency ≥10Hz, meeting the monitoring requirements for transient changes in pore water pressure (such as water inrush conditions); 2) Strain gauge: resolution ≤1με, operating temperature -20~80℃, suitable for the strain characteristics of surrounding rock and lining materials; 3) Seepage flow monitoring instrument: measurement range 0-100L / h, accuracy ±2%, adopts electromagnetic induction principle to reduce the impact of fluid disturbance.

[0025] II. Sensor Deployment and Implementation Steps

[0026] (1) The installation position calibration adopts the three-dimensional coordinate positioning method, and the spatial coordinates of each sensor are obtained by a total station. And it is matched with the underground structure BIM model to meet the following requirements: ,in, Design the coordinates of monitoring points in the BIM model.

[0027] (2) Sensor fixing and protection

[0028] 1) The strain gauges are bonded to the lining surface with epoxy resin to ensure a bond strength of ≥2MPa and avoid slippage error; 2) The water pressure sensor is buried in the surrounding rock through a borehole at a depth of ≥1m. It is connected to the pore water by permeable stone and a water-stop ring is installed on the outside to prevent interference from surface water.

[0029] III. High-Frequency Data Acquisition and Transmission Control

[0030] (1) Sampling frequency synchronization control: Each sensor node has a built-in crystal oscillator clock, which is uniformly synchronized through the main controller to ensure the stability of the 10Hz sampling frequency. Sampling interval s, actual sampling deviation Must meet: s; (2) The data transmission protocol adopts LoRaWAN low-power wide area network technology, and transmits data in packets every 30 seconds (each packet contains 300 sampling points), with a transmission delay of satisfy: To ensure the timeliness of data.

[0031] IV. Time Synchronization and Spatiotemporal Consistency Calibration

[0032] (1) Time synchronization is achieved by using GPS time synchronization + NTP protocol for dual synchronization, and the master node sets the reference time. Timestamps of each sensor node Must meet: ms, millisecond-level synchronization precision, synchronization period of 1 minute, and the node clock deviation correction value is updated after each synchronization. ; (2) Spatiotemporal consistency verification formula for multi-source data of the same monitoring section, such as pore water pressure With strain Its spatiotemporal matching degree C must satisfy: Where n is the number of sampling points, the closer C is to 1, the stronger the spatiotemporal correlation of the data, satisfying the physical coupling relationship between hydraulic load and structural response.

[0033] Through the above steps and quantitative control, high-frequency and synchronous monitoring of parameters related to hydraulic loads on underground structures is achieved by a distributed sensor network, providing spatiotemporally consistent multi-source data input for subsequent APNN models.

[0034] Furthermore, the data preprocessing includes: Outlier detection: Interquartile range (IQR) is used to identify outliers in pore water pressure, strain and other data, and values ​​exceeding Q1-1.5IQR or Q3+1.5IQR are removed. Missing value handling: Geological zoning interpolation method is used to fill in missing values ​​in offline sensor data; Standardization: Z-score standardization eliminates dimensional differences. The formula is as follows: ,in The mean, The standard deviation is denoted as .

[0035] Furthermore, the improved APNN model includes: Input layer: Receives static geological indicators and dynamic monitoring indicators; Model layer: Embedding physical constraint terms, the positive correlation between hydraulic load and pore water pressure is incorporated into the Gaussian function, and the corrected output is: Where λ is the physical constraint coefficient (0 < λ ≤ 1); Adaptive layer: When a change in pore water pressure >5% / h or a sudden change in surrounding rock strain is detected, the mean vector m of the model layer is automatically updated. i and standard deviation σ i .

[0036] Furthermore, the adaptive layer also includes: The model input is monitored in real time, and the input data of the indicator system is sampled at high frequency so that the model can perceive changes in the hydraulic load state of the underground structure in a timely manner. When a significant change in dynamic indicators is detected, an adaptive adjustment mechanism is triggered. The model automatically adjusts the parameters according to the real-time data to adapt to the new data distribution and characteristics, and performs online learning of the model to continuously adapt to the dynamic changes in the hydraulic load state of the underground structure and the evolution of the hydraulic load state of the underground structure.

[0037] This embodiment introduces physical constraints of hydraulic loads to correct the output function of the model layer and solve the physical irrationality of the pure data-driven model; it designs an adaptive triggering mechanism based on geological parameter mutations (such as pore water pressure gradient threshold) to replace the traditional data distribution change triggering method, which is more in line with the characteristics of underground engineering; a static-dynamic index fusion layer is added to the network structure to optimize the weight allocation of multi-source data and improve the generalization ability under complex geological conditions.

[0038] In a specific implementation, the improved APNN model further includes: For the calculation of hydraulic loads on underground structures, this paper improves the APNN adaptive probabilistic neural network by combining a teacher-student model framework, in-depth historical data mining, and data transfer from similar domains. This enhances the model's generalization ability and physical consistency under complex geological conditions. The technical solutions for improving the APNN model include: (1) Teacher-student model architecture design 1) Teacher Model Construction: Based on high-precision numerical simulation methods, such as finite element coupled analysis, input multi-source data: field measured pore water pressure, surrounding rock mechanical parameters, and physical simulation strain data, output hydraulic load benchmark values ​​containing physical constraints, such as the positive correlation between hydraulic load and pore water pressure, as the "knowledge output" of the teacher model.

[0039] 2) Student Model Optimization: Using an improved APNN as the student model, the output distribution of the teacher model is learned through knowledge distillation. A physical constraint term from the teacher model is introduced into the pattern layer, and the Gaussian function is modified as follows: ,in, These are the weighting coefficients. The physical constraint function output by the teacher model ensures that the prediction results of the student model conform to the evolution law of hydraulic load.

[0040] (2) Historical data augmentation strategy

[0041] 1) Based on multi-source historical data, such as field-measured pore water pressure time history data, physical simulation multi-condition strain data, and analytical calculation results, a "spatiotemporal correlation dataset" is constructed. Sample pairs are generated using the sliding window method, including: input features: porosity, water pressure change rate; output label: hydraulic load increment, thus expanding the training sample size.

[0042] 2) Based on the learning mechanism of APNN for historical failure cases, extreme conditions in historical data, such as sudden increases in pore water pressure, are weighted and sampled to improve the model's predictive sensitivity to sudden change scenarios.

[0043] (3) Transfer learning of data from similar domains

[0044] Approximate domain data, such as hydraulic load monitoring data from tunnel engineering under similar geological conditions, is introduced, and the data distribution is aligned using a domain adaptive layer. A new domain feature mapping module is added to the input layer of the APNN to transform geological parameters of the approximate data, such as rock integrity coefficients, into equivalent features of the target scene, such as porosity correction values, and the transfer weights are adjusted through cross-validation.

[0045] Based on the dynamic adjustment mechanism of APNN, when the distribution difference between the approximate data and the target data exceeds a threshold (such as KL divergence > 0.1), the network parameters are adaptively updated to avoid negative transfer.

[0046] (4) Optimization of model training process

[0047] Phase 1: The student model is pre-trained using historical data, and errors are corrected using physical constraints output by the teacher model. For example, when the student model's predicted value violates the rule "hydraulic load ≤ hydrostatic pressure", backpropagation is forced to correct the standard deviation of the model layer. .

[0048] Phase 2: Introduce approximate domain data for transfer training, optimize the input layer mapping parameters through a domain-adaptive loss function (such as adversarial loss), and verify the model's generalization ability by combining cross-validation methods.

[0049] Phase 3: Real-time input of newly collected field data, such as pore water pressure monitoring values, triggers the APNN adaptive mechanism to dynamically update the mode layer mean vector. This enables online learning.

[0050] In a specific embodiment, the steps to improve the APNN model are as follows: I. Teacher-Student Model Architecture and Design (1) Teacher model construction: The teacher model is constructed based on high-precision analytical methods and numerical simulation results, and the hydraulic load benchmark values ​​containing physical constraints are output.

[0051] 1) Teacher model input: Feature vectors from multiple data sources (Porosity, pore water pressure, elastic modulus of surrounding rock, Poisson's ratio of surrounding rock, elastic modulus of lining, Poisson's ratio of lining).

[0052] 2) Teacher model output: Hydraulic load benchmark values ​​that satisfy physical laws Its constraints are: and ,in, Let ρ be the density of water, g be the acceleration due to gravity, and h be the head height, based on the principle that hydraulic load is proportional to pore water pressure and less than hydrostatic pressure.

[0053] (2) Student Model (Improved APNN) and Distillation Loss Function The student model takes the improved APNN as its core and learns the physical constraints and output distribution of the teacher model through knowledge distillation.

[0054] 1) Student model output: APNN predicted value Its mode layer output is corrected as follows: ,in, The physical constraint weights (0 < λ ≤ 1). and are the mean vector and standard deviation of the i-th training sample, respectively, based on the pattern layer parameters borrowed from APNN.

[0055] 2) Distillation loss function: Combining the teacher model output and physical constraints, the formula is: ,in, This is the KL divergence (a measure of distributional dissimilarity). The mean square error between the predicted and measured values. and The first term is the weighting coefficient, and the third term is the physical constraint penalty term (to ensure that the predicted value does not exceed the hydrostatic pressure).

[0056] II. Historical Data Processing and Training Steps

[0057] (1) Historical data preprocessing is based on multi-source historical data (field measurements, physical simulations, and analytical results), and the following data preprocessing methods are adopted: 1) Outlier Detection: Outliers are identified using the interquartile range (IQR). The formula is: Outlier Judgment: or , It is the first quartile. It is the third quartile. .

[0058] 2) Standardization: Z-score standardization is used, with the following formula: , The mean, The standard deviation is denoted as .

[0059] (2) Weighted training with historical data: For extreme working conditions, such as historical data of sudden rise in pore water pressure, weighted sampling is used to improve the sensitivity of the model. 1) Sample weights Calculation formula: , Let be the pore water pressure of the k-th sample. The value represents the average pore water pressure; a larger value indicates a more extreme sample.

[0060] 2) Training objective: Minimize the weighted loss function .

[0061] III. Steps for Transfer Learning from Similar Domain Data

[0062] (1) Definition and mapping of approximate domain data Approximate domain data refers to hydraulic load data of underground engineering under similar geological conditions, such as measured data of adjacent tunnels, which are transformed into target scene data through feature mapping: 1) Mapping function: This maps the geological parameters of the approximate data. Convert to equivalent parameters of the target scene The formula is: M is the mapping matrix, b is the bias term, and data calibration is performed through physical simulation.

[0063] (2) The domain adaptive constraint borrows the adaptive mechanism of APNN, and triggers parameter updates when the domain difference exceeds a threshold: 1) Domain Disparity Measurement: The distributional distance between the target data and approximate data is calculated using the Maximum Mean Difference (MMD). , This is the feature mapping function.

[0064] 2) Adaptive triggering condition: If , If the preset threshold is used, then the APNN mode layer parameters are updated. , This is for adjusting the coefficient.

[0065] IV. Model Training Process

[0066] Phase 1: Pre-train the student model based on historical data, and correct the loss function using the physical constraints output by the teacher model (100-200 iterations). Phase 2: Introduce approximate domain data and fine-tune the mapping matrix M and APNN parameters using domain adaptive loss (50-100 iterations). Phase 3: Real-time input of new monitoring data triggers the APNN adaptive mechanism for dynamic updates. and This enables online learning.

[0067] Through the above steps, the improved APNN model combines the physical constraints of the teacher model, in-depth mining of historical data, and the ability to transfer data from similar domains, reducing the prediction error by 15%-20% while ensuring computational efficiency.

[0068] Furthermore, the improved APNN model training includes: dividing 70% of the data into a training set and 30% into a test set, and using dynamic learning rate gradient descent to optimize the parameters.

[0069] Furthermore, it also includes model validation, which includes verifying mean squared error and physical plausibility.

[0070] In a specific implementation, a numerical calculation system for the hydraulic load of underground structures based on artificial intelligence and multi-source data is provided, such as... Figure 2 As shown, it includes: Data Acquisition Module: Used for multi-source data acquisition of underground structures and construction of an indicator system; Preprocessing module: used to preprocess the multi-source data to obtain a standard dataset; Model building module: Used to build and train improved APNN models using standard datasets; Prediction module: Used to input real-time data into the improved APNN model to obtain numerical prediction results of hydraulic loads on underground structures.

[0071] In a specific embodiment, taking a deep-buried tunnel (350m deep, moderately weathered sandstone strata) as an example, the numerical calculation method for underground structure hydraulic load based on artificial intelligence and multi-source data is specifically explained, including: I. Implementation of Multi-Source Data Acquisition Indicator System Construction Static geological indicators: Geological surveys have determined that the rock type of the tunnel is moderately weathered sandstone. Laboratory tests show that the initial porosity is 8.2%. The lining uses C30 shotcrete with an elastic modulus of 31.5 GPa and a Poisson's ratio of 0.2.

[0072] Dynamic monitoring indicators: Relying on the established distributed sensor network, equipment was deployed at key sections of the tunnel from K12+350 to K12+450. Twenty water pressure sensors were installed on the outer edge of the lining, 15 sets of strain gauges were deployed in the deep rock (5m), and three seepage flow monitoring instruments were set up in each construction section. 10Hz high-frequency sampling was adopted, and GPS time synchronization technology (error <5ms) was used to ensure the spatiotemporal consistency of the data.

[0073] II. Data Preprocessing Implementation

[0074] Outlier detection: IQR analysis was performed on the pore water pressure data from June 12, 2024. The calculated values ​​were Q1=0.32MPa, Q3=0.58MPa, and IQR=0.26MPa. Seventeen outliers exceeding 0.32-1.5×0.26=-0.07MPa or 0.58+1.5×0.26=0.97MPa were removed.

[0075] Missing value handling: For the data from the strain gauge at K12+380 after 2 hours of offline operation, the geological zoning interpolation method was used to fill in 3600 missing strain values ​​based on the linear variation trend of three adjacent monitoring points.

[0076] Standardization: The seepage flow data were standardized using Z-fraction. The mean value of this parameter is 12.5 L / min and the standard deviation is 3.2 L / min. The measured value of 18.3 L / min at a certain moment was processed to obtain (18.3-12.5) / 3.2=1.81.

[0077] III. Improving the Application of APNN Models

[0078] Model parameter settings: The physical constraint coefficient λ is set to 0.8, and the adaptive trigger threshold is set to 5% / h of pore water pressure change rate. The input layer contains 6 features (rock type encoding, initial porosity, lining elastic modulus, real-time pore water pressure, surrounding rock strain, and seepage flow).

[0079] Model layer correction: When the pore water pressure is 0.45MPa, the original Gaussian function output is 0.62. After introducing physical constraints, it is corrected to 0.62×(1+0.8×0.45 / 0.5)=0.62×1.72=1.067 (assuming the hydrostatic pressure reference value is 0.5MPa).

[0080] Adaptive adjustment example: On June 15, 2024, at 14:23, a sudden increase in pore water pressure was detected from 0.42 MPa to 0.65 MPa (change rate 54.8% / h), triggering the adaptive mechanism. The model layer mean vector m... i Updated from [0.41, 280, 3.2] to [0.63, 310, 4.1], standard deviation σ i Adjustments will be made accordingly.

[0081] IV. Teacher-Student Model Training

[0082] Teacher model output: Based on finite element analysis, when the pore water pressure is 0.5MPa, the output hydraulic load reference value is 0.48MPa (satisfying the hydrostatic pressure constraint of ≤0.5MPa).

[0083] Knowledge distillation process: The student model initially predicted 0.52 MPa, which was corrected by the distillation loss function. The KL divergence term makes the distribution approximate the teacher model, and the physical penalty term (0.52-0.5=0.02) forces backpropagation, and finally corrected to 0.49 MPa.

[0084] V. On-site application results

[0085] During the heavy rain in July 2024, the system successfully captured three transient hydraulic load changes. Among them, at 16:45 on July 18, the pore water pressure suddenly increased (from 0.38 MPa to 0.72 MPa within 10 minutes). The improved APNN model completed the parameter adaptive update within 1.2 seconds, and the predicted hydraulic load was adjusted from 0.36 MPa to 0.68 MPa. The relative error with the synchronously acquired measured value of lining pressure (0.69 MPa) was only 1.4%, which is a significant improvement in accuracy compared with the traditional APNN model (error 8.7%).

[0086] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0087] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A numerical calculation method for hydraulic loads on underground structures based on artificial intelligence and multi-source data, characterized in that, include: Conduct multi-source data collection on underground structures and construct an indicator system; The multi-source data is preprocessed to obtain a standard dataset; An improved APNN model was built and trained using a standard dataset; The real-time data is input into the improved APNN model to obtain the numerical prediction results of the hydraulic load on the underground structure.

2. The numerical calculation method for hydraulic loads of underground structures based on artificial intelligence and multi-source data according to claim 1, characterized in that, The indicator system includes static geological indicators and dynamic monitoring indicators.

3. The numerical calculation method for hydraulic loads of underground structures based on artificial intelligence and multi-source data according to claim 2, characterized in that, The static geological indicators include: rock type, initial porosity and lining material parameters, and the dynamic monitoring indicators include real-time pore water pressure, surrounding rock strain and seepage flow.

4. The numerical calculation method for hydraulic loads of underground structures based on artificial intelligence and multi-source data according to claim 1, characterized in that, The data preprocessing includes: Outlier detection: Outliers are identified using the interquartile range method; Missing value handling: Geological zoning interpolation method is used to fill in missing values ​​in offline sensor data; Standardization: Eliminate dimensional differences through Z-score standardization.

5. The numerical calculation method for hydraulic loads of underground structures based on artificial intelligence and multi-source data according to claim 1, characterized in that, The improved APNN model includes: Input layer: Receives static geological indicators and dynamic monitoring indicators; Model layer: Embedding physical constraint terms, the positive correlation between hydraulic load and pore water pressure is incorporated into the Gaussian function, and the corrected output is: Where λ is the physical constraint coefficient; Adaptive layer: When the pore water pressure change rate > k, the mean vector m of the model layer is automatically updated. i and standard deviation σ i .

6. The numerical calculation method for hydraulic loads of underground structures based on artificial intelligence and multi-source data according to claim 5, characterized in that, The adaptive layer also includes: The model input is monitored in real time, and the input data of the indicator system is sampled at high frequency so that the model can perceive changes in the hydraulic load state of the underground structure in a timely manner. When a significant change in dynamic indicators is detected, an adaptive adjustment mechanism is triggered. The model automatically adjusts the parameters according to the real-time data to adapt to the new data distribution and characteristics, and performs online learning of the model to continuously adapt to the dynamic changes in the hydraulic load state of the underground structure and the evolution of the hydraulic load state of the underground structure.

7. The numerical calculation method for hydraulic loads of underground structures based on artificial intelligence and multi-source data according to claim 1, characterized in that, The improved APNN model includes: dividing 70% of the data into a training set and 30% into a test set, and using dynamic learning rate gradient descent to optimize the parameters.

8. The numerical calculation method for hydraulic loads of underground structures based on artificial intelligence and multi-source data according to claim 1, characterized in that, It also includes model validation, which includes checking the mean square error and physical plausibility.

9. A numerical calculation system for hydraulic loads on underground structures based on artificial intelligence and multi-source data, characterized in that, include: Data Acquisition Module: Used for multi-source data acquisition of underground structures and construction of an indicator system; Preprocessing module: used to preprocess the multi-source data to obtain a standard dataset; Model building module: Used to build and train improved APNN models using standard datasets; Prediction module: Used to input real-time data into the improved APNN model to obtain numerical prediction results of hydraulic loads on underground structures.

Images